跳轉到

Project Lifecycle Guide

This guide defines the standardized phases of a research project at AGILAB, from "Initiation" to "Archive."

1. Initiation Phase

  1. Create Repo: Initialize a new repository based on SoftwareTemplate.
  2. Set up W&B: Create a corresponding project in Weights & Biases.
  3. Define Data Storage: Determine if large datasets should be stored on server /data paths or external storage.

2. Development Phase

  • Atomic Commits: Habitually git commit with descriptive messages.
  • Sync Protocol: Notify team members whenever environment.yml changes.
  • Regular Testing: Ensure pytest passes at all times.

3. Pre-publication Phase

Before submitting a paper or releasing a preprint: 1. Code Cleanup: Remove all print debugging statements; use logging instead. 2. Documentation Completion: Ensure README has complete installation and execution examples. 3. Set License: Default recommendation is MIT License or Apache-2.0. 4. Version Tagging: Tag the current code (e.g., git tag v1.0-submission).

4. Archive Phase

After paper publication: 1. Server Weight Backup: * Move final model weights (Best Checkpoints) to the unified lab storage path (e.g., /data/checkpoints/). * Naming Convention: [ProjectName]_[Date]_[KeyMetric].pt (e.g., Humanoid_20240501_reward300.pt). 2. Path Documentation: Record the server hostname (or ID) and the absolute path to these weights in the project's README under the "Archive" section (e.g., Server: G108-A6000, Path: /data/checkpoints/...). 3. Handover Docs: If there are successors, write a simple INTERNAL_HANDOVER.md in docs/ explaining specific machine paths or hardware interfaces.